Skip to content

Notes · · 4 min read

Why a Manufacturing AI Should Ask Which Model Before It Answers

A maintenance technician asks the plant's AI assistant for the supply pump flow on a cleaning skid. The reply comes back in a second: 25 cubic metres an hour, stated clearly and confidently. The technician sets up the job around that number.

The skid on the floor is the larger model. Its pump is rated for 35.

The assistant gave the right answer for the wrong machine.

Illustrative walkthrough using an original demo manual. For demo purposes only.

One manual, two machines

Equipment manuals are usually written for a family of models. The manual in this example covers a clean-in-place skid sold in two versions: the S, with one tank, and the D, with two. Both share the same document, the same pump tags and the same specification table.

On page 8, Table 2-1 lists the supply pump P-201 for each model:

ItemCIP-400 SCIP-400 D
Supply pump P-20125 m³/h at 4.0 bar g35 m³/h at 4.5 bar g
Electrical supply400 V, 63 A400 V, 100 A
Safety valve PSV-2016.0 bar g6.0 bar g

The values come from an original demo manual written for testing. For demo purposes only.

Same row, same page, two correct answers. Which one is right depends entirely on the machine in front of the person asking.

Experienced engineers handle this without thinking. They check the nameplate first. A reliable assistant needs the same habit.

Why a guessed value is a risk

When an assistant picks one column, it is wrong some of the time and looks right every time. The wrong number arrives with the same tone, the same formatting and the same confidence as a correct one. Nothing on the screen signals that a choice was made.

In a plant, that number becomes a pump setting, a breaker size or a spare part order. The cost of a confident guess lands downstream, with the person who trusted it.

What a good assistant does

A well-built assistant follows one simple rule. When the manual gives a value separately by model and the user hasn't said which model applies, it asks one short question, names the options exactly as the manual prints them, and waits.

Which model are you working on: CIP-400 S or CIP-400 D?

Once the user answers, the assistant gives the value for that model and shows the page it came from. The technician can open the source and see the exact line in the table.

Three details make this work in practice:

  • It uses the manual's own names. The user picks from the model numbers printed in the document, which match the nameplate.
  • It keeps one model per answer. Values from different models are never blended into one reply.
  • It leaves the choice with the user. The person at the machine knows which model it is. The assistant asks them.

Asking only when it matters

Every question costs the operator time, so a good assistant asks only when the manual requires it.

The safety valve PSV-201 is set to 6.0 bar g on both models. There is one value, so the assistant answers directly and cites the page.

The rule is easy to state. When the manual splits a value by model, the assistant asks. When the manual gives one value, it answers.

An instruction is a starting point. A test is proof.

This behavior comes from a written instruction to the AI model, combined with ordinary code that first decides which manual the question is about. An instruction describes what the assistant should do. Testing shows what it actually does.

The dependable approach is to write a test question for every place the manual splits a value by model, leave the model out of each question, and record the result. A pass means the assistant asked. A single unlabeled number means it guessed, and that case gets fixed before the team relies on it.

Check your own assistant this afternoon

You can run this check on any assistant your team already uses.

  1. Find a value your manuals list separately by model. Pump flow, supply voltage, a torque or a setpoint all work well.
  2. Ask for it without naming the model.
  3. Read the reply.
    • It asks which model: pass.
    • It gives both values, each clearly labeled: acceptable.
    • It gives one number: it guessed. That is worth knowing before anyone acts on it.
  4. Open the source. Whatever the reply says, follow it to the page. If you can see the page, you can check the number.

A good assistant knows when the right answer is a question

The most useful thing an assistant can say is sometimes "which one?" That question is the habit your best engineers already have, carried into the tool your whole team uses.

If you'd like to see this tested on your own manuals, the Assistant Pilot runs it on one department's documents and gives you the measured results. If your team already has an assistant, the workshop shows you how to test it together.

By Virginia Kalesić. All notes.

What does your plant need to know when your best person is out?

Start with one department and the questions that keep coming up. We'll talk through the documents you have and whether a fixed-price build is a good fit.

Book a 30-minute call